# Resting ECG Heart Rate Task Card Task id: `resting_hr` Evidence status: grounded Evidence reviewed: 2026-08-15 ## Target `resting_hr` predicts `ecg.hr_bpm`: heart rate in beats per minute from the 10-second resting 12-lead ECG acquired during the HPP clinical visit. Implementation facts: - Task: `RestingHrTask` - target dataset: `ecg` - target field: `hr_bpm` - target source: None - target unit: bpm - default research stage: `00_00_visit` - valid range: 30.0-130.0 bpm; a PhenoBench plausibility filter, not a clinical interval - primary metric: `r2` - Additional scalar metrics: `rmse`, `mae`, `pearson_r`, `spearman_r` when target and prediction are non-constant. ## Why This Matters Diament 2023 used sleep-derived heart-rate / HRV features to predict resting heart rate measured by 10-second ECG at the testing center. The paper reported that sleep heart-rate features improved MAE from 7.13 bpm for age/sex/BMI to 5.19 bpm, with correlation 0.69. The PhenoBench task represents the stable target. Paper-inspired configs attach specific sleep-derived feature surfaces. **What a row here does not support.** The target is a **clinic** resting heart rate from a 10-second ECG at one visit. It is not a resting heart rate in the physiological sense of a stable trait, not an overnight or ambulatory measure, and not a cardiovascular risk estimate. A single 10-second acquisition in a clinical setting carries white-coat and posture effects that an overnight measurement does not. **The setting is the whole subtlety on this Task, and the corpus already records it.** Heart rate is a WatchPAT input channel, so a sleep or WatchPAT arm can look like it is recomputing the target when it is not. The `cgm_sleep_ecg` configs state the resolution in their `review.source_match_rationale`: the target is a clinic resting heart rate read from the `ecg` dataset while a WatchPAT arm reads an overnight `watchpat_hr` channel, so two heart-rate measurements in two settings make a high score **cross-setting transfer rather than derivation**. Anyone adding a heart-rate-bearing modality here should declare the same way rather than leave a reader to infer it. ## Related HPP Measurements Demonstrated in HPP and persisted in this repository, from an internal review record, test split: - V1 sleep + demographics reaches test R2 = 0.454 (95% CI [0.410, 0.495]) on a cohort of 9,382. - V1 Nightingale + demographics reaches R2 = 0.127 (95% CI [0.094, 0.159]) on 9,420. The sleep result is the largest single-modality effect in the cardiovascular family of cards, and it is the direction Diament et al. reported - overnight heart-rate features carry most of the recoverable signal in a clinic ECG heart rate. The gap to Nightingale metabolomics is large, but the two tracks are **not** the same people: 9,382 against 9,420 participants under different `cohort_id` hashes. Similar sizes do not establish common support, so read this as an unmatched corpus comparison rather than a controlled contrast. **Against the demographic floor both tracks clear**: sleep **ΔR2 = +0.384 (95% CI [+0.343, +0.424])**, Nightingale **+0.058 ([+0.034, +0.086])**. The sleep figure is a within-track margin, and it is the informative one: almost the whole of that track's raw 0.454 is added signal rather than floor, which is what makes the setting caveat load-bearing rather than pedantic. The two tracks sit on different cohorts (9,382 against 9,420), so sleep's larger margin is not a paired comparison against Nightingale. Read it with that caveat: part of that margin is two measurements of the same physiological quantity in different settings, which is why the `benchmark_claim` on those arms is transfer and not derivation. **These are `reviewable` rows, not `final` ones.** Measured across the committed leaderboards: all 787 rows carry `curation_state: reviewable` and none is `final`. A complete fresh v0 re-run is committed to, so every figure in this section is pre-re-run evidence that the re-run supersedes - persisted and citable today, not a frozen result. **Which curation states the Figure 1 export accepts is the export track's live question, not this card's**; read the two counts above, not a gate. ## Clinically Meaningful Variants **A categorical variant is not defensible on a single 10-second clinic acquisition.** Bradycardia and tachycardia are defined at 60 and 100 bpm, which are the closest things to fixed thresholds anywhere in this card set - but they are clinical findings interpreted with symptoms, medication and context, not labels applied to one research measurement. Three specific blocks: - **Medication.** Beta blockers and rate-controlling drugs move heart rate directly, so a bradycardia bin would largely label treated participants. The Task does not model medication. - **Fitness inverts the meaning.** A trained endurance athlete below 60 bpm is the healthy end of the distribution while a bradycardic patient is not, and nothing in the target distinguishes them. - **Acquisition.** A 10-second clinic ECG with white-coat effect is not the resting condition the thresholds assume; the overnight channel discussed above would be closer, and it is a different measurement. The continuous target stays primary. If a rhythm-abnormality Task is wanted, it should come from a rhythm label rather than from a bin on this scalar. ## Baselines And Ceilings The floor is a no-modality-feature baseline: `NoFeaturesPredictor` contributes no modality features, and `DemographicFloorStrategy` fits age, sex, and BMI. The Diament-inspired comparison config is: ```text the internal evaluation configuration ``` It uses the stable-loader WatchPAT heart-rate summary fields: - `heart_rate_mean_during_sleep` - `heart_rate_min_during_sleep` - `heart_rate_max_during_sleep` - `heart_rate_mean_during_nrem` - `heart_rate_mean_during_rem` - `heart_rate_mean_during_wake` These fields are averaged across baseline sleep nights before fitting. The natural ceiling is the paper's broader HRV feature surface, not the current stable-loader heart-rate summary row. Any sleep-feature claim should compare against the demographic floor on matched support and state whether it is a stable-loader analogue or a paper-faithful HRV route. ## Physiological Time-Series V1 the internal evaluation configuration evaluates a multi-rate residual temporal encoder trained end-to-end on target-blind actigraph, PAT-infrared, SpO2, and heart-rate windows from up to three baseline WatchPAT nights. It retains the same clinic ECG target and matched demographic floor. It remains paper-inspired because it does not reproduce the paper's full 447-feature HRV surface. This route is a target-adjacent pipeline positive control because nocturnal heart rate is an input. On the exact 860-participant test support, the temporal model reached R² 0.4408 and MAE 5.114 bpm. A train-only Ridge using sampled HR mean/SD plus age/sex/BMI reached R² 0.4519 and MAE 5.198 bpm. Temporal minus proxy delta-R² was -0.0111 (95% CI [-0.0400, +0.0164]); delta-MAE was -0.084 bpm (95% CI [-0.227, +0.062]). This supports ingestion/training validity, not a temporal-representation advantage. ## Caveats This is not an exact Diament HRV replication. The paper feature group mainly uses the broader HRV feature surface, including the `sleep_all_hrv` table and the paper's 447-feature matrix. The stable PhenoBench config uses loader-exposed heart-rate summaries only. The exact HRV feature surface should become either a promoted loader key (`sleep_hrv`) or an artifact-backed Predictor before claiming paper-faithful replication. Beating demographics is insufficient for this task. Every model claim must also compare with a sampled-HR proxy on identical support. ECG `hr_bpm` is a short resting measurement. It is affected by recent activity, clinic context, medications, caffeine, arrhythmia, and measurement time. It is a useful autonomic-tone target, not a complete cardiovascular endpoint. ## Metrics Primary metric is R2. Secondary MAE in bpm is clinically interpretable and is the easiest comparison to the Diament paper. Pearson and Spearman are useful rank/linear diagnostics when target and prediction are non-constant. ## References - Diament et al. 2023, "A Multimodal Dataset of 21,412 Recorded Nights for Sleep and Respiratory Research" - HPP documentation paper-task run: `research_runs/phenobench-paper-task/diament_sleep_hr_2026_06_18/` ## Benchmark-Track Evidence Benchmark-Track Evidence reviewed: 2026-07-29. V1 Nightingale and sleep are proposed beyond age/sex/BMI. Sleep heart rate is deliberately retained as a target-adjacent positive control and must be compared with the sampled-HR proxy; ECG inputs are excluded. See `04-cardiovascular-lipid.md`. ## v0 Final-Test Status The classical release is complete: Ridge/logistic and GBDT are remotely verified on every approved exact cohort and rendered in the Figure 1 task deep dives. ## Open Questions - Should loader-native HRV features be promoted so this can become a paper-faithful Diament route? - Which medications or arrhythmia flags should be excluded or audited? - Should clinic ECG resting HR be paired with sleep nocturnal-HR-dipping tasks?